Open financial LLMs, as a research resource
FinGPT is a collection of open financial large language models and the recipes to build them, from the AI4Finance Foundation. Its core idea is that financial data shifts quickly, so instead of training a model from scratch the way an expensive proprietary effort would, you adapt an open base model to finance cheaply with LoRA. What you actually get is a set of Jupyter notebooks, published model weights on Hugging Face, and datasets covering tasks like financial sentiment analysis and named-entity work. At 20,502 stars as of 2026-06, it is a well-known entry point for financial NLP.
The framing to keep straight: FinGPT is a research and teaching resource, not a product you plug in and run.
What it is
The repository centers on tasks rather than a single model. The most mature and documented piece is financial sentiment analysis, with instruction-tuned models you load as a LoRA adapter on top of an open base model. There are also multi-task and forecasting experiments, datasets for instruction tuning, and notebooks that walk from data preparation through fine-tuning to inference. It sits alongside sibling projects in the AI4Finance Foundation such as FinRL for reinforcement-learning trading.
How to use it
The common path is loading a published sentiment model as a PEFT adapter:
from transformers import AutoModelForCausalLM
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-13b-hf")
model = PeftModel.from_pretrained(base, "FinGPT/fingpt-sentiment_llama2-13b_lora")Fine-tuning your own adapter is the deeper path and is documented in the notebooks, though it needs a capable GPU. Pin your dependency versions (transformers, peft, bitsandbytes), because version drift is the most common reason a notebook stops working.
Read this: it is not a trading system
This matters before you build expectations. FinGPT provides language models for financial text; it does not place trades, manage a portfolio, or constitute investment advice, and the project states as much. The forecasting demos are research artifacts, their accuracy is not validated with a built-in backtest, and at least one hosted demo has gone offline due to funding. So FinGPT is a source of components, like a sentiment signal you might feed into a larger system, not a system that decides or trades on its own.
Where it fits, and where it doesn’t
Reach for FinGPT when you want a financial sentiment or NLP model to use as a feature, when you want to learn how LoRA fine-tuning for a domain works, or when you want open weights and datasets to build on. As a research resource it is genuinely useful, and the MIT license is permissive.
It is the wrong thing to expect a turnkey trading or forecasting product from. Reproducibility is the main friction: notebooks lag behind fast-moving dependencies and break, some datasets and the Chinese-market pieces lack maintained data sources, and the forecasting results are not something to trust as predictions. For a trading framework or a quant platform, the alternatives below fit better.
FinGPT versus the trading and quant projects
| FinGPT | qlib | TradingAgents | ai-hedge-fund | |
|---|---|---|---|---|
| Stars | 20,502 | 44,469 | 85,821 | 60,129 |
| Type | financial LLM resources | quant research platform | multi-agent trading framework | educational agent simulator |
| Provides | models, datasets, notebooks | data, factors, models, backtest | LLM trading agents | investor-persona agents |
| Trading | no | backtest and execution | execution-capable | no, simulation |
| Best for | financial NLP, fine-tuning | quant ML research | agent strategies | learning multi-agent design |
Counts are from GitHub as of June 2026. qlib is the full ML research platform. TradingAgents and ai-hedge-fund are LLM-agent trading approaches. FinGPT’s niche is the language-model layer, which you would combine with one of the others rather than use alone for trading.
The star curve
The star curve rose with the surge of interest in domain-specific LLMs, when “a financial GPT” was a compelling idea and the cheap-fine-tuning pitch resonated. The growth reflects research and educational interest more than production deployment, which fits what the repository actually is.
What the issue tracker warns you about
The recurring pain is reproducibility. Notebooks break as transformers, peft, and bitsandbytes evolve, so you often need to debug imports and pin versions. Model loading errors, especially around quantization and LoRA merging, are common. Datasets and the Chinese-market forecasting pieces have missing or unmaintained data sources. Fine-tuning is GPU-heavy and the original training data for some models was not released, so exact reproduction is hard. And a hosted forecasting demo went offline due to funding, which signals where maintenance energy is limited.
Related
For a quant research platform, see qlib. For agent-based trading, see TradingAgents and ai-hedge-fund. For the broader LLM tooling landscape, see trending LLM tooling. For what is climbing, see the daily digest and the weekly report.
FAQ
Can I use FinGPT to trade? No. FinGPT provides financial language models, not a trading system. It does not place trades or give investment advice, and the forecasting demos are unvalidated research artifacts. Use it as a component, like a sentiment signal, inside a larger system you build and validate.
What does FinGPT actually give me? Open financial language models (notably for sentiment analysis), datasets for instruction tuning, and notebooks covering data preparation, LoRA fine-tuning, and inference. It is a research resource, not a packaged application.
How do I use a FinGPT model? Load a published LoRA adapter on top of an open base model with the PEFT library. Pin compatible versions of transformers, peft, and bitsandbytes, since version drift is the most common cause of failures.
FinGPT vs qlib, what is the difference? FinGPT is the language-model layer for financial text. qlib is a full quant ML research platform with data, factors, models, and backtesting. They are complementary: you might use a FinGPT sentiment signal as a feature inside a qlib pipeline.
Is FinGPT hard to reproduce? Often, yes. The notebooks lag behind fast-moving dependencies and break, some datasets lack maintained sources, and original training data for some models was not released. Pin versions and expect to debug; favor the more recent, maintained model variants.